---
license: apache-2.0
base_model:
- Qwen/Qwen3.6-35B-A3B
- nightmedia/Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx
- Hcompany/Holo3-35B-A3B
- samuelcardillo/Qwopus-MoE-35B-A3B
- Qwen/Qwen-AgentWorld-35B-A3B
- llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved
tags:
- gguf
- llama.cpp
- qwen3_5_moe
- moe
- merge
- vision-language
- security
- image-text-to-text
---

# RazorStrike-v1 GGUF

GGUF build of `lancejames221b/razorstrike-v1`. **Same lineage as the current MLX repo**, not the older DARE-TIES build documented in `lancejames221b/razorstrike-v1-bf16` (that lineage was superseded 2026-07-21 and has no GGUF/MLX quant under this org).

**Base**: a 4-bit-equivalent quantization of `nightmedia/Qwen3.6-35B-A3B-Holo3-Qwopus-AgentWorld-qx64-hi-mlx` (itself a merge of `Hcompany/Holo3-35B-A3B`, `samuelcardillo/Qwopus-MoE-35B-A3B`, and `llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved` with `Qwen/Qwen-AgentWorld-35B-A3B`) with the Qwen3.6-35B-A3B vision tower spliced back in — see `lancejames221b/razorstrike-v1`'s README for the full provenance and known-issues notes (including the inherited repetition/looping caveat).

## Files

- `razorstrike-v1-IQ4_XS.gguf` — 4.37 BPW IQ4_XS text model, converted with llama.cpp from a raw-HF-compatible reconstruction of the source above.
- `mmproj.gguf` — matching multimodal projector for image input.
- `RAZORSTRIKE_V1_GGUF_MANIFEST.json` — build and verification notes.

The bf16 GGUF intermediate was generated and smoke-tested locally but is not uploaded because it is ~65 GiB.

## Fix in this build

Previous GGUF attempts generated corrupted text because the source checkpoint was MLX-derived. MLX stores non-linear-attention text RMSNorm weights with the `+1.0` offset already baked in, while the llama.cpp Qwen converter expects raw HF-style weights and applies its own `+1.0` transform for those tensors. This build reconstructs the HF-naming checkpoint with:

- `model.language_model.*norm.weight` shifted by `-1.0`, except `linear_attn.norm.weight`.
- `linear_attn.norm.weight` kept unchanged.
- vision tower norms kept unchanged.
- MLX `switch_mlp.{gate,up}_proj.weight` merged into raw HF `experts.gate_up_proj`.
- MLX `switch_mlp.down_proj.weight` renamed to raw HF `experts.down_proj`.
- llama.cpp conversion run with `--no-mtp` to avoid false extra-layer metadata.

## Local verification

Verified on the fixed IQ4_XS GGUF through `llama-server`:

```bash
llama-server \
  -m razorstrike-v1-IQ4_XS.gguf \
  --mmproj mmproj.gguf \
  -c 4096 \
  -fit off
```

Text smoke test:

```text
System: <|think_off|>
User: What is 17 times 24? Answer directly.
Assistant: 408
```

Image smoke test with a generated PNG containing a green square, yellow circle, and `TEST-42`:

```text
Assistant:
- Green square
- Yellow circle
- Text: "TEST-42"
```

## License

Apache-2.0, matching the Qwen3.6 lineage and the current RazorStrike-v1 model card.